基于非参数独立分量分析的说话人识别方法

Speaker Recognition Based on Nonparametric Independent Component Analysis

首先用非参数独立分量分析方法提取表征说话人音频特性的时域基函数组,语音信号可由这些基函数线性组合而成。每个可识别的说话人对应一个不同的基函数组,对某个特定人的输入音频,只有与它对应的基函数组使其系数向量各分量之间的独立性最强(也就是互信息最小)。对待识别音频,分别用已知说话人的时域基函数组计算各自的系数向量,并计算系数向量各分量之间的互信息。互信息最小的基函数组对应的说话人即为识别结果。实验结果表明,即使用很少的测试数据.也能达到很高的识别率。

Time-domain basis functions are obtained through nonparametric independent component analysis first,whicn exhibit the main characteristics of the specific speaker. Speech signals then can be represented by the superposition of the basis functions. Every speaker candidate has his own set of basis functions, which are different from those of others. And,for a speech signal by a specific speaker,only his own set of basis functions can make the elements of the coefficient vectors most independent （namely, the mutudl information is minimal）. To recognize a test speech signal,all sets of basis functions are used to produe the coefficient vectors,and the mutual information among the elements of the coefficient vectors are calculated. The speaker who has the minimum mutual information is thought of as the producer of the test speech signal. Experiments show that a high recognition rate can be achieved by a small amount of data.